Mutational-signature adjusted models reveal HLA-mediated depletion of antigenic mutations in non-small cell lung cancer (NSCLC)

Background : Quantifying immune editing in human tumors is challenging because observed mutation patterns reflect both mutational processes and selection. Whether HLA class 1 genotypes uniformly influence which somatic mutations persist remains unclear. Methods : Whole-genome sequencing (WGS) from CPTAC NSCLC tumors with matched blood and adjacent normal lung underwent germline single nucleotide polymorphism (SNP), somatic single nucleotide variant (SNV), HLA typing, neoepitope prediction, and mutational signature decomposition using established pipelines. SNP/SNVs were annotated with codon position, resulting amino acids, and evolutionary similarity metrics (BLOSUM62, PAM250, Miyata, Atchley, Grantham). Four-digit HLA alleles were collapsed into recognized supertypes; neoepitopes with IC50 ≤150 nM were considered high-affinity. Models of expected nonsynonymous mutations were created using the GRCh38 reference, mutational signature, and tumor mutation burden and compared to observed SNP/SNVs using cosine similarity. Statistical significance was assessed using ANOVA and generalized linear models with FDR correction. Results : 219 cases (111 LUAD, 108 LUSC) were processed successfully. Median SNPs were 862.3, SNVs were 103.8 with an associated median of 62.5 predicted neoepitopes (4.9 high-affinity) per case. Amino acid substitutions arising from the third codon position (N=96) had higher similarity than those from position 1 and 2 (N=226), (p<0.001). SBS4 dominated 167 (76.3%) tumors; in blood, SBS5 and SBS58 characterized 152 (69.4%) and 55 (25.1%) cases, respectively. Codon-position biases differed across signatures, with SBS4 showing relative depletion of position-3 substitutions compared to SBS5 and SBS58 (p<0.001). Signature-based modeling demonstrated cosine similarity >0.90 with higher cosine similarity in SBS4 compared to SBS5 and SBS58. ANOVA comparisons of amino acid substitutions revealed multiple supertypes (A02, B27, B44) with relative depletion of anchor substitutions and high-affinity neoepitopes with these substitutions (FDR<0.05). Conclusion : Germline SNPs exhibit strong evolutionary constraint, whereas somatic SNVs show minimal intrinsic codon-position bias, supporting the hypothesis that cancer mutagenesis is a random process. Controlling for mutational signature uncovers HLA-specific immune editing, demonstrating that antigenic mutations are selectively depleted in specific HLA contexts. This signature-conditioned framework provides a clinically actionable platform for improving neoantigen prediction and immuno-oncology translation.